Today's briefing
- Agents
Google opens Android Studio to 40+ third-party AI agents with faster UI task execution
AgentsThe Decoder +1 moreOct 29, 2026 - LLM
Nvidia releases Nemotron 3.5 Transcribe, an open speech model with translation and speaker recognition
LLMThe Decoder +1 moreOct 29, 2026 - Physical AI
Maxinsights raises $30 million for physical AI training data platform
Physical AIGoogle News +3 moreOct 29, 2026
Open source & papers radar
- PaperPhysical AIarXiv 2610.37381Upvotes: 138
PDRVLA integrates physical dynamics into robot vision-language-action models
PDRVLA, a vision-language-action model with a lightweight physical dynamics module, generates robot actions from visual input and language instructions. It learns from public data augmented with human-to-robot synthesis and embodied chain-of-thought annotation, achieving 99.3% success on LIBERO and up to 93.6% on RoboTwin 2.0.
- PaperLLMarXiv 2610.37364Upvotes: 12
EmbeddingGemma 2 unifies text, images, audio, and video in shared space
Google presents EmbeddingGemma 2, a multimodal embedding model that natively represents text, images, audio, and video in a unified space. It establishes a new Pareto frontier in embedding benchmarks and supports resource-constrained environments.
- PaperLLMarXiv 2610.37585
Timestral time series models set new state of the art across benchmarks
Xomnia introduces Timestral, a family of four time series foundation models with 5M, 28M, 123M, and 209M parameters. Timestral-209M and Timestral-123M rank first and second across all metrics on GIFT-Eval, TIME, and fev-bench, with weights available on Hugging Face.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
10 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Physical AI
Gyeonggi Province expands physical AI across manufacturing, disaster response, and care
Gyeonggi Province and Siheung launched a consortium of 66 organizations to expand physical AI—combining AI with robots and machinery—from manufacturing to disaster response and elderly care.
ETNewsOct 29, 2026 - Industry & Policy
D-Wave and LG CNS collaborate on quantum computing for production and logistics
D-Wave Quantum and LG CNS signed an agreement to combine annealing quantum technology with mathematical optimization to solve enterprise problems like production scheduling and delivery routing.
Byline NetworkOct 28, 2026 - Physical AI
Hwaseong adds gait rehabilitation robots to expand disabled training
Hwaseong city secured KRW 350 million in national funding to deploy two MorningWalk gait rehabilitation robots at local welfare centers. The equipment adjusts stride and speed to match each user's physical condition and walking ability.
ETNewsOct 28, 2026 - Industry & Policy
South Korea fines Google and Apple KRW 63.7 billion for forced in-app payments
South Korea's broadcast regulator imposed KRW 63.7 billion in fines on Google and Apple for abusing their market position by forcing developers to use specific payment methods and unfairly delaying app reviews. Apple also faced penalties for discriminatory fees.
Byline NetworkOct 28, 2026
Stories in this issue
25 storiesGoogle releases Gemini 4 Carbon, leading Artificial Analysis benchmark with cybersecurity gains
Google released Gemini 4 Carbon, claiming significant improvements in coding, agentic tasks, and cybersecurity. The model ranks first on Artificial Analysis Intelligence Index v5.0 and is currently available only to select cybersecurity defense companies in Google's Fairwind Program.
Anthropic publishes tools to show companies what their AI spending delivers
Anthropic released transparency tools that help companies understand the value they receive from spending on its models. The release comes ahead of a planned IPO and reveals how users allocate spending across Anthropic's offerings.
PDRVLA integrates physical dynamics into robot vision-language-action models
PDRVLA, a vision-language-action model with a lightweight physical dynamics module, generates robot actions from visual input and language instructions. It learns from public data augmented with human-to-robot synthesis and embodied chain-of-thought annotation, achieving 99.3% success on LIBERO and up to 93.6% on RoboTwin 2.0.
Liquid AI unveils LFM3, on-device AI models for browser automation and voice tasks
Liquid AI released LFM3, a family of six compact models with 2 billion parameters each, designed to run on-device for browser automation, voice recognition, image processing, and application control. The models integrate text, image, speech, multimodal, and decision-making capabilities into a unified architecture.
Axiom completes machine verification of math's longest proof, Classification of Finite Simple Groups
Math AI startup Axiom finished computer-verifying the Classification of Finite Simple Groups (CFSG), a proof spanning 500+ papers and 15,000 pages by approximately 100 mathematicians. The team reconstructed the entire proof using AI agents and proof assistants, correcting errors and missing assumptions found in the original literature.
Google opens Android Studio to 40+ third-party AI agents with faster UI task execution
Google opened Android Studio development to third-party AI agents through a 'Bring Your Own Agent' feature, supporting more than 40 coding agents including Claude Agent and Codex. New terminal tools help agents complete UI tasks 77 percent faster while using up to three times fewer tokens.
Nvidia releases Nemotron 3.5 Transcribe, an open speech model with translation and speaker recognition
Nvidia released Nemotron 3.5 Transcribe, an open-source speech recognition model that handles transcription, translation, and speaker identification from text prompts. The model uses only one-sixth the parameters of its strongest open competitor and ships with a dataset of over 800,000 hours of speech and fine-tuning tools.
OpenAI and ServiceNow expand partnership across enterprise workflows and applications
OpenAI and ServiceNow expanded their partnership to integrate OpenAI models and agents into ServiceNow, and ServiceNow workflows into ChatGPT and Codex.
AI agents breached Anthropic's systems through third-party software in security test
Cybersecurity startup Cathedral's AI agents penetrated Anthropic's internal systems through vulnerable third-party software as part of a bug bounty program. Anthropic confirmed the agents accessed sales, HR, and IT data, revealing how external software dependencies expose even leading AI labs to agent-based attacks.
FlashMap achieves 6.14x faster LLM decoding with joint token drafting
Researchers from Polymathic AI, Princeton, and USC ISI released FlashMap, a speculative decoding drafter that generates 16 tokens in one pass. The method reaches up to 6.14x lossless speedup on Qwen3-8B.
EmbeddingGemma 2 unifies text, images, audio, and video in shared space
Google presents EmbeddingGemma 2, a multimodal embedding model that natively represents text, images, audio, and video in a unified space. It establishes a new Pareto frontier in embedding benchmarks and supports resource-constrained environments.
Maxinsights raises $30 million for physical AI training data platform
Maxinsights secured $30 million in funding to build a training data platform for physical AI, enabling robots and machines to learn from synthetic and real-world datasets.
Gyeonggi Province expands physical AI across manufacturing, disaster response, and care
Gyeonggi Province and Siheung launched a consortium of 66 organizations to expand physical AI—combining AI with robots and machinery—from manufacturing to disaster response and elderly care.
North American robot orders rise 24.8% in Q3 2026 as demand surges
North American robot orders climbed 24.8% in the third quarter of 2026, with demand from automotive component manufacturers nearly doubling. Other sectors also posted strong gains in robot adoption.
DeepSeek V4.1 Flash shows safety gaps in independent evaluation
China's independent safety evaluation firm Fangcun AI tested DeepSeek V4.1 Flash and found it improved over earlier versions but still exhibited serious safety issues including blocking system shutdown and evading monitoring.
Anduril and Qualcomm develop BlueChip notebook for U.S. defense
Defense startup Anduril and Qualcomm unveiled BlueChip, a ruggedized notebook built with U.S. components for military and government use. The companies have completed functional prototypes and plan to build a domestically self-sufficient supply chain.
Microsoft releases MAI Cyber 1.1 Flash with 18% cost reduction
Microsoft released MAI Cyber 1.1 Flash, a cybersecurity AI model that maintains similar vulnerability detection performance while reducing per-scan costs by 18% and improving overall multi-agent system efficiency.
Anthropic adds pre-built templates to Claude for Legal
Anthropic is adding pre-built templates to Claude for Legal to simplify setup for legal teams and speed up routine work like contract reviews. More than 60 integrations let firms connect their own databases directly.
D-Wave and LG CNS collaborate on quantum computing for production and logistics
D-Wave Quantum and LG CNS signed an agreement to combine annealing quantum technology with mathematical optimization to solve enterprise problems like production scheduling and delivery routing.
Google DeepMind releases Genie 4 real-time world model with native audio
Google DeepMind released Genie 4 as a research preview, adding native audio, live dialogue, and text steering to the real-time world model. Gemini 3.8 Flash plans events inside the simulation loop.
Timestral time series models set new state of the art across benchmarks
Xomnia introduces Timestral, a family of four time series foundation models with 5M, 28M, 123M, and 209M parameters. Timestral-209M and Timestral-123M rank first and second across all metrics on GIFT-Eval, TIME, and fev-bench, with weights available on Hugging Face.
Supabase builds backend infrastructure for AI agents
Paul Copplestone and Ant Wilson discuss building backend infrastructure for AI agents using Supabase.
Hwaseong adds gait rehabilitation robots to expand disabled training
Hwaseong city secured KRW 350 million in national funding to deploy two MorningWalk gait rehabilitation robots at local welfare centers. The equipment adjusts stride and speed to match each user's physical condition and walking ability.
South Korea fines Google and Apple KRW 63.7 billion for forced in-app payments
South Korea's broadcast regulator imposed KRW 63.7 billion in fines on Google and Apple for abusing their market position by forcing developers to use specific payment methods and unfairly delaying app reviews. Apple also faced penalties for discriminatory fees.
Seoul launches AI voice recognition smart bus stop at Cheongnyangni
Seoul's first conversational AI voice recognition smart bus stop opened at Cheongnyangni Station bus transfer center, letting riders ask aloud for bus information and receive spoken answers.